By HIROSHI GOTO
Is your problem rooted in your high-speed technology design or in the equipment you're using to test it?
The importance of test-signal fidelity grows as data rates increase. That's because waveform inadequacies cause one of two business problems: One, they make component flaws seem worse than they really are and increase costs by forcing unnecessary redesigns; two, and more troublesome, they may conceal flaws and allow products with performance deficiencies to be shipped. Bit-error-rate testers (BERTs) such as the one shown in Figure 1 cast the final judgment of component and system performance -- BER -- making the selection of the proper BERT imperative to successful design and production of high-speed devices and systems.
Evaluating channels and receivers begins with the test waveform. For accurate compliance and useful diagnostic testing, pattern generators must have sufficient bandwidth to excite at least three harmonics. Figure 2 shows an instrument-grade eye diagram that has smooth, continuous edges with no harmonic distortion, bandwidth limiting, overshoot, or ringing, even at 25 Gbps. BERTs that produce such eye diagrams can support all three harmonics.
Test waveforms free of ISI and DCD
To achieve eye diagrams like the one shown in Figure 2, test waveforms must have minimal inter-symbol interference (ISI) and duty-cycle distortion (DCD). ISI, the dominant signal impairment at high data rates, is caused by the frequency response of the channel. It gets its name because the frequency content associated with a given symbol depends on the values of neighboring symbols. Long strings of identical symbols carry lower frequencies, and rapidly alternating strings carry higher frequencies. Pre-/de-emphasis at the transmitter and equalization at the receiver are techniques for solving ISI problems.
|FIGURE 1. BERTs are essential tools for characterizing the performance of communications components, subsystems, and systems.|
Simulations can predict the effects of a channel on a waveform at the earliest design stages. Of course, simulation accuracy depends on channel characterization accuracy. Simulations have trouble with real world edge effects caused by connectors, vias, dielectric anisotropies, etc., and have to be verified by measurement.
Extracting ISI from a simulation enables starting with an ideal waveform and examining it at any point in its propagation. The benefit of a measurement -- provided the engineer starts with an instrument-quality waveform and takes care with any cables and connections -- is that the engineer sees reality. Internally generated ISI is a big problem for engineers trying to make such accurate measurements.
Any intrinsic ISI in the test waveform corrupts measurements of channel ISI. It's possible to unfold the ISI of the initial waveform from the ISI of the outgoing waveform. But extensive, detailed analysis is required.
DCD is the variation in the widths of positive and negative pulses, i.e., differences in the widths of 1s and 0s. It can be identified by vertical asymmetries in eye-diagram crossing points of alternating clock-like data patterns. DCD introduces extra frequency content that alters measurements of channel ISI and is difficult to untangle.
|FIGURE 2. Today, BERTs must provide eye diagrams with enough resolution to capture all three harmonics at data rates of 25 Gbps or more.|
Quality pattern generators provide adjustable crossing points for a variety of reasons, including the ability to remove DCD. When selecting a pattern generator, examine the transmitted waveform and make certain that neither memory-based patterns nor pseudo-random binary sequence (PRBS) patterns exhibit ISI or DCD.
De-emphasis opens high-data-rate eyes
Since ISI is caused by a channel's frequency response, its effects can be mitigated by transmitting a signal that exhibits the inverse of that response. That's the idea behind pre- and de-emphasis.
For example, the first approximation of channel-frequency response is a low-pass filter. To mitigate this low-pass nature, high-frequency signal components are amplified or low-frequency components are attenuated. Pre-emphasis amplifies high frequencies and increases the total signal power, while de-emphasis attenuates the lower frequencies and signal power either decreases or remains constant. Since the effects on the signal are equivalent up to an overall constant, both can be referred to as "emphasis."
Taps and cursors
High-frequency components of a digital signal occur at logic transitions. The first order correction applies a larger voltage swing to the bit that follows a data transition and is called a "two-tap" correction. In principle, this process can continue with finer corrections until the complete inverse frequency response of the channel is encoded in the transmitted signal. In practice, we're limited to applying correction factors to individual bits or else face the increased cost of requiring a transmitter with much greater bandwidth than the signal itself requires.
Each tap is a corrective boost or reduction applied to the amplitude of a bit neighboring the bit of interest. The bits to which these taps are applied are called "cursors." The terminology for bits that precede those of interest is a "pre-cursor" and for bits that follow, "post-cursor."
Since tap values reflect the inverse frequency response of the channel, their optimum values can be calculated from that frequency response. The most accurate way to characterize a channel's response is to measure its S-parameters with a vector network analyzer. To determine the optimal taps, transmission analysis software can be used with the BERT and a "four-tap" emphasis module.
To assure BER measurements accurately reflect the performance of the devices being tested, it's imperative that the pattern generator have minimal random jitter (RJ). A 1-psec RMS error in RJ becomes a 14-psec (peak-to-peak) mistake in total jitter (TJ) for a 10-12 error ratio. At 25 Gbps, 3-psec RMS of RJ is sufficient to cause a >10-12 BER. At extreme data rates, the intrinsic RJ of an instrument-quality waveform must be <350-fsec RMS to prevent the test equipment from contributing more than 10% of uncertainty to BER measurements.
De-embedding -- the process of subtracting transmission-line effects from measurements -- can help limit jitter. Effective de-embedding has two strict requirements. First, the S-parameters of any channel (including cables, PCB traces, and the connections between them) must be scrupulously characterized to bandwidths of at least 3X the data rate. Second, no matter how accurate the characterization of the network elements being de-embedded, the accuracy of the waveform extracted at the position of interest is still limited by the quality of the test waveform. Any ISI or DCD intrinsic to the test waveform must also be de-embedded, an arduous task that cannot be performed with S-parameters and simulation alone.
|FIGURE 3. There are two kinds of crosstalk: synchronized (a) and unsynchronized (b).|
Effects of crosstalk
To create systems at 100 Gbps and higher, the advantages of high-speed serial technology -- differential signaling, embedded clocking, transmitter emphasis, and receiver equalization -- are combined with the data-rate scaling properties of parallel architectures. For example, 100 Gigabit Ethernet achieves 100 Gbps by combining either 10 separate differential parallel lanes, each at 10 Gbps, or four differential lanes, each at 25 Gbps. The problem of skew is solved by adding complexity to the data-link layer of the protocol stack. Crosstalk -- the other major problem of parallel architectures -- can't be corrected so easily.
There are two categories of crosstalk. Synchronized crosstalk occurs in systems where the separate lanes operate with the same clock. Since the victim and aggressors are frequency locked, crosstalk noise on the victim has fixed phase relationships (see Figure 3a). Unsynchronized crosstalk occurs when the victim and aggressors operate with distinct clocks. Here, the timing of aggressor noise varies over the victim eye diagram (see Figure 3b).
Designers can make headway against crosstalk through simulation, but just as is the case for ISI, the system ultimately has to be tested with parallel streams of data. Measuring the BER effects of crosstalk requires that the pattern generator have special features. It must be able to transmit several simultaneous signals. At 10--20 Gbps, three parallel signals are usually sufficient. At higher rates, where the unexpected must be anticipated, it may be necessary to excite every lane in the system.
The first step toward evaluating channel crosstalk is to assure that the multichannel pattern generator itself doesn't already have crosstalk. An easy way to check is to perform a BERT loopback test comparing bathtub plot measurements between single-channel and multichannel operation.
Crosstalk BER measurements can't be performed on an oscilloscope. While oscilloscopes can separate different types of jitter and estimate BER performance by extrapolation, they come up short with crosstalk. In general, TJ can only be measured on a BERT, and in the case of crosstalk, TJ can't even be reliably estimated on an oscilloscope. Simulation can provide insight into what level of crosstalk to expect; if the channels are synchronized, simulation also can help engineers identify bathtub plot asymmetries. But the magnitude of the crosstalk impact on BER can only be measured.
Beyond the cutting edge
Good test engineers eliminate systematic errors caused by worn-out cables, dirty connections, and poorly trained technicians. But systematic uncertainty caused by test equipment can be invisible right up until a system fails.
Engineers working at the cutting edge of technology must use test tools just beyond cutting edge. Accurate channel and receiver testing begins with the quality, calibration, and reliability of the pattern generator.
Tight budgets may make it tempting to use inexpensive equipment in a lab. Yet cutting corners by using flawed pattern generators with poorly integrated error counters will increase the net cost of both development and production. Inaccurate product characterization costs more in market share than the difference between a high-end BERT that produces high-quality waveforms and a less expensive alternative that generates inadequate waveforms.
HIROSHI GOTO is Anritsu Co.'s business development manager for general purpose test equipment. He has more than 25 years of experience in communications. He began his career in design engineering of optical products and has spent the past 10 years in product marketing for telecom and datacom products.
Past Lightwave Articles